Codex vs Claude Code (2026): which one fits your work?
Codex is OpenAI's coding agent and comes with ChatGPT plans. Claude Code, Anthropic's coding agent, comes with paid Claude plans. Both work in the terminal, code editors, desktop and mobile apps, and on cloud computers their vendor runs. They differ in the models they run, how usage is billed, and whether you can read the agent's source.
Superconductor publishes this page. Superconductor runs both Codex and Claude Code as agent harnesses and is not affiliated with OpenAI or Anthropic.
Codex vs Claude Code at a glance
Each capability below is marked Built in, Partial, or Not offered for each product.
| Capability | Codex | Claude Code |
|---|---|---|
| Run agents on your own computer | Built in. On your machine, the Codex CLI and the ChatGPT desktop app run the agent locally. Codex Cloud runs tasks on OpenAI's machines. | Built in. Local sessions run on your own machine. As of November 2026, self-hosted environments for cloud sessions are in beta on Team and Enterprise. |
| Edit the code yourself mid-task | Built in. Inside VS Code, Cursor, Windsurf, JetBrains, and Xcode, you edit the same files the agent is changing. | Built in. Local sessions edit the files you have open, and teleport pulls a cloud session into your terminal. |
| Read and change the agent's source code | Built in. The Codex CLI is open source under the Apache-2.0 license. | Not offered. As of November 2026, the source is proprietary under Anthropic's Commercial Terms. |
| Sandbox the agent's commands on your computer by default | Built in. Commands run in an OS-enforced sandbox that writes only to the workspace. Network access is off until you allow it. | Partial. By default, an auto-mode classifier approves or blocks commands instead of asking. The sandboxed Bash tool is opt-in. |
| Pick up instructions written for the other agent | Partial. AGENTS.md is the instruction file. CLAUDE.md is read only after you add it as a fallback name in the configuration. | Partial. CLAUDE.md is the instruction file. AGENTS.md is read in a project that has no CLAUDE.md. |
| Run the agent from scripts and CI | Built in. The codex exec command runs a task without the interactive interface. The Codex GitHub Action runs a task in CI. | Built in. The claude -p command runs a task without the interactive interface. The Agent SDK for Python and TypeScript embeds Claude Code. |
| Let teammates take part in an agent's session | Not offered. On a Mac, you can share a read-only snapshot of a local Codex thread with workspace members. Nobody else can continue a shared thread. | Partial. Teammates can open a shared cloud session link. As of November 2026, Claude Tag is in public beta on Team and Enterprise and takes requests from anyone in a Slack channel. |
| Start agents automatically from alerts and events | Partial. Scheduled tasks in Codex run on a timer. Gmail, Slack, and GitHub event triggers run as ChatGPT scheduled tasks on the web and mobile. | Built in. Cloud scheduled tasks start a session on a schedule, on a GitHub event, or from an API call. |
Support for different types of teams
Each task is marked Can, Partly, or Cannot for each product.
| Task | Codex | Claude Code |
|---|---|---|
| Plan a feature from a specification or a planning call | Partly. In Plan mode, Codex drafts a plan before it edits code. Someone has to paste decisions from a meeting into the prompt. | Partly. Plan mode has Claude Code research the codebase and propose a plan before it edits anything. As of November 2026, Anthropic documents no feature that joins a planning call. |
| Implement a feature or bugfix | Can. A Codex Cloud task returns a diff you can open as a pull request. | Can. Sessions run on your own machine or on a cloud computer Anthropic operates. A cloud session pushes a branch, and you open the pull request. |
| Gather feedback on the implementation | Partly. A Codex Cloud task returns changed files and test results for teammates to read in ChatGPT or GitHub. | Partly. A shared cloud session link shows the session's latest state, without live updates. Teammates comment on the pull request in GitHub. |
| QA the implementation | Partly. Tests run in the task's cloud VM. Browser tests need Codex on your own computer, using the ChatGPT desktop app or a browser MCP server. | Partly. The agent runs the test suite inside the session. In the desktop app, a Browser pane lets it check the running app. |
| Review the code before it merges | Can. Pull requests in GitHub get a Codex review automatically or when someone comments @codex review. The review follows rules you write in AGENTS.md. | Can. Code Review, on Team and Enterprise plans, checks pull requests in the repositories an admin selects, when they open, on every push, or on request. |
| Deploy and monitor | Partly. Your existing pipeline handles deployment. The Codex GitHub Action runs Codex inside CI. | Partly. Deployment stays with your existing pipeline. Agents read production errors through Model Context Protocol (MCP) servers you connect. |
| Use your existing coding agent subscription plan | Partly. You pay with a ChatGPT plan, an OpenAI API key, or OpenAI models billed through Amazon Bedrock or Google Vertex AI. Claude subscriptions are not accepted. | Partly. A paid Claude plan, an Anthropic API key, or a Bedrock, Google Cloud, or Foundry account pays for Claude Code. Other providers' subscriptions do not work. |
| Task | Codex | Claude Code |
|---|---|---|
| Fix a bug even when no engineer is available | Partly. After an admin shares a cloud environment, workspace members can start a fix from Slack, and paid users from Linear. | Partly. Start a cloud session from the Claude app or Slack, then review the pull request on GitHub. Anthropic documents no running preview for cloud sessions. |
| Work from Slack or Microsoft Teams | Partly. An @ChatGPT mention in Slack can start a Codex Cloud task. @ChatGPT in Microsoft Teams has no documented Codex handoff. | Partly. Tag @Claude in Slack to start a coding task. Claude in Microsoft Teams has a waitlist as of November 2026. |
| Work on the go | Can. The ChatGPT iPhone app starts Codex Cloud tasks, and the iPhone and Android apps can steer Codex on your Mac or Windows PC. | Can. The Claude iPhone and Android apps start and monitor cloud sessions. Desktop apps run on macOS and Windows, with Linux in beta as of November 2026. |
| Answer a question or produce a report without touching the codebase | Can. Besides code, Codex in the ChatGPT desktop app works on documents and spreadsheets. Plugins and MCP servers connect it to other apps. | Partly. Sessions work in a repository or a local folder and reach other tools through MCP. Claude Cowork, included in paid Claude plans, handles tasks outside code. |
| Keep your bill predictable | Partly. At the plan's Codex limit, new work waits for the reset. Plus, Pro, and flexible-pricing Business workspaces can buy ChatGPT credits, and anyone can use an API key. | Can. Pro, Max, Ultra, and Team plans have a fixed monthly price and usage limits. Turning on usage credits bills work past the limits at API rates. |
| Avoid depending on a single model provider | Partly. A ChatGPT plan covers only OpenAI's models, although the CLI can also run local open models. | Cannot. Anthropic supports only Claude models in Claude Code. |
| Replace ChatGPT or Claude as your team's AI assistant | Can. A ChatGPT plan includes Codex. One plan and one desktop app handle chat, office documents, and coding. | Can. Every paid Claude plan includes Claude Code and the Claude chat app. Anyone with a paid seat can use the chat app as a general assistant. |
| Task | Codex | Claude Code |
|---|---|---|
| Keep agents inside the network perimeter | Partly. Each Codex Cloud environment allows all public internet access, or only package managers and the domains you list. Tailscale can reach private services. | Can. Each cloud environment allows no access, a trusted allow-list, full access, or a custom list of domains. |
| Pass a security review | Can. OpenAI's SOC 2 Type II report covers ChatGPT Business and Enterprise. Codex follows the workspace's access, retention, and residency settings. | Can. Anthropic lists SOC 2 Type II, SSO, SCIM, audit logs, and a Compliance API, and Enterprise has a HIPAA-ready offering. |
| Use your existing version control and project management systems | Partly. Codex Cloud tasks work with GitHub, and a GitLab beta handles merge request reviews and @codex tasks. Linear is built in, and Jira connects through Atlassian's MCP server. | Partly. Anthropic-hosted cloud sessions open pull requests on GitHub only. Self-hosted environments and local sessions can work with other Git hosts. Linear and Jira connect through MCP. |
| Deploy and manage company-wide rollouts | Can. ChatGPT Enterprise has custom roles, SCIM groups, and separate switches for local and cloud Codex. Admins enforce settings with a managed requirements file. | Can. The Team plan covers 2 to 150 users with standard and premium seats. Admins push server-managed settings and can deploy a company-wide CLAUDE.md to every machine. |
| Control spend | Can. Enterprise admins set monthly per-user credit limits and workspace-wide spend controls. Business admins manage credits and spend controls for the workspace. | Can. Team and Enterprise admins set spend controls per user and per organization. Usage analytics show who uses Claude Code. |
How Codex and Claude Code work
Codex runs in the terminal, on the web, and in the ChatGPT desktop and iPhone apps. Editors get it through an extension for VS Code, Cursor, and Windsurf, or through the JetBrains and Xcode integrations, and the ChatGPT Android app can steer Codex on a connected computer.
On your own computer, Codex edits files in your working copy. In Codex Cloud, each task runs in its own VM on OpenAI's machines, starts from an environment prepared from your repositories, and can make up to four attempts at once. The task returns changed files and test results, which you commit or open as a pull request. Cloud tasks start from ChatGPT, an @ChatGPT mention in Slack, Linear, a GitHub comment, or the CLI.
Claude Code runs in the terminal, in VS Code and JetBrains editors, in the Claude desktop app, on the web at claude.ai/code, and in the Claude mobile app. A local session edits files on your computer. A cloud session runs on a computer Anthropic operates, where it clones the repository from GitHub and pushes its changes to a branch. From there you open a pull request, or teleport the session into your terminal and keep working.
Models and who pays for tokens
Codex runs OpenAI's models. As of November 2026, OpenAI recommends GPT-6.1 Sol, GPT-6 Astra, GPT-6.1 Terra, and GPT-6 Luna, and you set a reasoning level for each task. A ChatGPT plan covers that usage up to its limits. With an OpenAI API key, you pay per token instead. The Codex CLI can also reach OpenAI models through Amazon Bedrock or Google Vertex AI, or run open models locally through Ollama or LM Studio.
Claude Code runs only Anthropic's models. As of November 2026, those are Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5, and Anthropic lists a 1M-token context window for each. A paid Claude plan covers that usage up to its limits, and an Anthropic API key moves you to per-token pricing. Enterprises can also run Claude Code on Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry.
Codex vs Claude Code pricing (ChatGPT plans vs Claude plans)
OpenAI and Anthropic each include their coding agent in a monthly chat plan. Claude Code draws on the same limits as Claude chat. Codex draws on the same limits as ChatGPT Work, OpenAI's agent for documents and multistep tasks.
Codex
OpenAI includes Codex in ChatGPT plans with usage limits. On Plus, Pro, and flexible-pricing Business workspaces, ChatGPT credits pay for usage past them. A ChatGPT plan's Codex allowance depends on the plan and the model. As of November 2026, the Plus allowance resets every five hours and weekly limits can also apply, while ChatGPT Pro plans have no five-hour limit. Codex shares the allowance with ChatGPT Work, OpenAI's agent for documents and other multistep work. Developers who would rather pay per token sign in with an OpenAI API key.
- Free. Codex in the ChatGPT desktop app with GPT-6 Luna, subject to rollout. No CLI, IDE extension, or Codex Cloud.
- Go, $8 a month. The same desktop-app access as Free.
- Plus, $20 a month. Adds the Codex CLI, IDE extension, Codex Cloud, automatic GitHub code review, and Slack.
- Pro, from $100 a month. Tiers at $100, $200, and $500 a month with larger allowances.
- Business, $25 per user a month. $20 per user a month billed annually. Adds a shared workspace, admin controls, and ChatGPT credits for usage past each seat's limit.
- Enterprise, custom. SCIM, role-based access control, data residency, and Compliance API audit logs.
Codex pricing figures come from OpenAI's Codex pricing page as checked on November 4, 2026. Current pricing is available at learn.chatgpt.com/docs/pricing.
Claude Code
Anthropic charges a monthly Claude plan per person. Claude Code shares the plan's usage limits with the Claude chat app. Pro and Max usage resets on a rolling 5-hour window and also has a weekly limit. Verified Ultra accounts have only the weekly limit. Beyond Claude plan usage limits, you can enable usage credits at standard API rates or use a Console account billed per API token. A Claude plan covers Claude Code's tokens and Anthropic's cloud sessions. You cannot use the API credits included with Claude Max, Ultra, and Team in Claude Code.
- Pro, $20 a month, or $17 a month billed annually. Claude Code in the terminal, editor, desktop, web, and mobile apps.
- Max, $100 a month. Five times Pro's usage per 5-hour window and per week.
- Ultra, $200 a month. Ten times Pro's weekly usage and no 5-hour limit for verified accounts.
- Team, $25 a seat a month, or $125 a premium seat ($20 and $100 billed annually). 2 to 150 users. A premium seat has five times a standard seat's usage. SSO, usage analytics, and spend controls.
- Enterprise, $20 a seat a month billed annually, plus usage at API rates. 20-seat minimum. SCIM, audit logs, the Compliance API, role-based access, IP allow-listing, and a HIPAA-ready offering.
Claude Code pricing figures match Anthropic's published rates as checked on November 4, 2026. Current pricing is available at claude.com/pricing.
Review and collaboration
Codex reviews GitHub pull requests on its own or when someone comments @codex review, following the review rules in AGENTS.md. A beta extends those reviews to GitLab merge requests.
On Team and Enterprise plans, Claude Code's Code Review checks pull requests automatically and follows the rules in CLAUDE.md and REVIEW.md. Claude Tag, a public beta on the same plans, lets anyone in a paired Slack workspace tag @Claude with a request.
Security and enterprise controls
ChatGPT Enterprise adds SCIM, role-based access, data residency, and Compliance API audit logs to Codex. Admins switch local and cloud Codex on separately, enforce settings with a managed requirements file, and can set monthly per-user credit limits.
Claude Enterprise adds role-based access and IP allowlisting to Claude Code. Turning on IP allowlisting blocks Anthropic-hosted cloud sessions, Code Review, and cloud scheduled tasks. Team and Enterprise admins set spend limits per person and per organization.
What our benchmark shows
Superconductor's benchmark takes merged pull requests from Superconductor's own Ruby on Rails codebase and infers a specification from each one. Each agent implements the specification without seeing the merged code, and evaluator models grade correctness, completeness, and code quality. Runs published since July 24, 2026 use 10 pull requests, one ticket each (Kimi K3 benchmark, August 3, 2026; Grok 4.6 benchmark, August 13, 2026).
Every result pairs one model with one harness. In the GPT-5.5 benchmark of April 24, 2026, GPT-5.5 at high effort in the Codex harness scored about 0.75 out of 1 at about $1.40 per ticket. Opus 4.7 in the Claude Code harness scored about 0.68 at about $9.50 per ticket. The July 24, 2026 audit removed runs that had found the merged solution and re-ran the benchmark with that access blocked. Afterward, GPT-5.4 at high effort in Codex and Fable 5 in Claude Code shared first place with a quality score of about 0.82 out of 1 on the post's chart.
No run has tested the same model in both harnesses, and every model tested is older than the current lineup. The results compare model and harness pairings on one codebase. They do not show which harness is better.
Who should pick Codex or Claude Code
Choose by which models you want and which plan already pays for them: OpenAI's GPT models under a ChatGPT plan (Codex), or Anthropic's Claude models under a Claude plan (Claude Code). You don't have to pick just one, though, and the next section explains why.
Who should pick Codex
- You already pay for ChatGPT and want one OpenAI plan to cover chat, documents, and coding.
- You want to read the agent's source code and have its local commands sandboxed with network access off by default.
- You want a Codex Cloud task to make up to four attempts at once so you can choose among the results.
Who should pick Claude Code
- You want Anthropic's Claude models, including a 1M-token context window on Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5.
- You want anyone in a Slack channel to hand Claude a request through Claude Tag, in public beta on Team and Enterprise.
- You want agents to start on a schedule, on a GitHub event, or from an API call without anyone opening a session.
Should you commit to one coding agent?
Probably not. The model that writes the best code for your team this month may not be the one that does next month. In April 2026, GPT-5.5 at high effort in Codex had the highest quality score on Superconductor's benchmark, ahead of Opus 4.7 in Claude Code. In the July re-run, Fable 5 in Claude Code tied GPT-5.4 at high effort in Codex for first. None of those runs tested today's models, so the next release could reorder them again.
Access can change too. Italy's privacy regulator blocked ChatGPT in March 2023 (Garante), and in early 2025 Italy blocked DeepSeek while NASA, the US Navy, and Texas barred it from government devices (TechCrunch). In February 2026, the US president ordered federal agencies to stop using Anthropic's technology, with six months for agencies such as the Department of Defense to phase it out (CBS News). And if all of your agent work runs through one vendor's plan, you have little room to push back when that vendor raises prices or cuts usage limits. The Multiplayer AI Manifesto, written by Superconductor's team, makes the same case for staying provider-agnostic.
Keeping both set up is cheap. Claude Code reads AGENTS.md when a project has no CLAUDE.md, so one instruction file serves both agents, and you can switch the week the results change.
Can you use both?
Yes. Superconductor, a multiplayer AI platform, runs Codex and Claude Code on the same ticket, so you can see which one handles your code better before you lean on either. Each agent works on its own cloud computer that runs your whole app, so you can compare the two diffs and live previews side by side. A connected ChatGPT Plus, Pro, Business, Enterprise, or Edu plan pays for Codex's tokens, and a connected Claude Pro, Max, or Team plan pays for Claude Code's. Superconductor plan credits cover the cloud computers. Superconductor runs the Codex and Claude Code CLIs itself and does not import Codex Cloud tasks or Claude Code cloud sessions.
Frequently asked questions
Is Codex better than Claude Code?
Neither is better; they fit different work. Codex runs OpenAI's GPT models from an open-source CLI, editors, and the ChatGPT apps, and a ChatGPT plan pays for it. Claude Code is limited to Anthropic's Claude models, which a paid Claude plan covers in the terminal, editors, and the Claude apps. As of November 2026, Superconductor's benchmark has not run the same model in both harnesses, so it cannot say which harness writes better code.
Can I use Claude Code and Codex together?
Yes. Claude Code reads AGENTS.md in a project that has no CLAUDE.md, so one instruction file can serve both agents. Superconductor, a multiplayer AI platform, runs Codex and Claude Code on the same ticket, each on its own cloud computer, and shows the two diffs and live previews side by side. A connected ChatGPT Plus, Pro, Business, Enterprise, or Edu plan pays for Codex's tokens, and a connected Claude Pro, Max, or Team plan pays for Claude Code's tokens.
Should I use both Claude Code and Codex?
Probably, if you can. In Superconductor's benchmark, GPT-5.5 at high effort in Codex had the highest quality score in April 2026, and in the July re-run Fable 5 in Claude Code tied GPT-5.4 at high effort in Codex for first, so the better agent for your code can change between releases. Keeping both also leaves you an option if one vendor raises prices, tightens usage limits, or loses access where you work.
How much do Codex and Claude Code cost in 2026?
As of November 4, 2026, ChatGPT Free and Go include Codex in the desktop app with GPT-6 Luna, subject to rollout. The Codex CLI and Codex Cloud start with ChatGPT Plus at $20 a month, ChatGPT Pro costs $100, $200, or $500 a month, and Business costs $25 per user a month. Claude Code comes with Claude Pro at $20 a month, Max at $100, Ultra at $200, and Team at $25 a seat a month ($125 for a premium seat).
Which models do Codex and Claude Code use?
As of November 2026, Codex uses OpenAI's GPT models, and OpenAI recommends GPT-6.1 Sol, GPT-6 Astra, GPT-6.1 Terra, and GPT-6 Luna. The Codex CLI can also run local open models. Claude Code uses only Anthropic's Claude models: Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5.
Claude Code vs Codex for large codebases
As of November 2026, Anthropic lists a 1M-token context window in Claude Code for Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5. OpenAI's Codex docs give GPT-6.1 Sol a 500,000-token default context window in Codex. Codex and Claude Code can each hand parts of a task to subagents.
Do Codex and Claude Code work in Slack and GitHub?
Yes, both do. Codex starts a cloud task from an @ChatGPT mention in Slack and reviews pull requests when someone comments @codex review in GitHub. Claude Code takes Slack requests through Claude Tag, in public beta on Team and Enterprise as of November 2026, and on those plans Code Review checks GitHub pull requests.
Is Codex or Claude Code open source?
The Codex CLI is open source under the Apache-2.0 license, and its code is on GitHub at openai/codex. Claude Code is proprietary software under Anthropic's Commercial Terms as of November 2026. Claude Code's GitHub repository holds the changelog and issue tracker rather than the source code.
Sources (43)
Codex's facts come from OpenAI's Codex documentation and the open-source Codex repository on GitHub. Claude Code's facts come from Anthropic's product, pricing, and help pages and the Claude Code changelog. Both were checked on November 4 or 5, 2026, and each source below gives its date. Benchmark figures come from Superconductor's dated blog posts.
- Codex docs: Pricing (checked November 4, 2026)
- Codex docs: Models (checked November 4, 2026)
- Codex docs: Cloud environments (checked November 4, 2026)
- Codex docs: IDE extension (checked November 4, 2026)
- Codex docs: ChatGPT desktop app (checked November 4, 2026)
- Codex docs: Codex on mobile (checked November 4, 2026)
- Codex docs: Agent approvals and security (checked November 4, 2026)
- Codex docs: AGENTS.md (checked November 4, 2026)
- Codex docs: Advanced configuration (model providers) (checked November 4, 2026)
- Codex docs: Non-interactive mode (checked November 4, 2026)
- Codex docs: GitHub Action (checked November 4, 2026)
- Codex docs: Review GitHub pull requests with Codex (checked November 4, 2026)
- Codex docs: Review GitLab merge requests with Codex (checked November 4, 2026)
- Codex docs: Use ChatGPT in Slack (checked November 4, 2026)
- Codex docs: Use ChatGPT (sharing a read-only snapshot) (checked November 4, 2026)
- Codex docs: Admin setup (checked November 4, 2026)
- openai/codex on GitHub (README, license, and model catalog) (checked November 4, 2026)
- Claude pricing (checked November 4, 2026)
- Claude Code for Enterprise (checked November 4, 2026)
- Claude Code docs index (checked November 4, 2026)
- Claude Code changelog (checked November 4, 2026)
- Claude Code license (checked November 4, 2026)
- Claude Help Center: Set up Code Review for Claude Code (checked November 4, 2026)
- Claude docs: Work with Claude Tag (checked November 4, 2026)
- Claude Help Center: Use Claude Code with your Pro, Max, or Ultra plan (checked November 4, 2026)
- Claude download page (checked November 4, 2026)
- Superconductor blog: GPT-5.5 is out, and it's the new quality-per-dollar leader (April 24, 2026) (checked November 4, 2026)
- Superconductor blog: Preventing agents from cheating on our custom SWE-Bench (July 24, 2026) (checked November 4, 2026)
- Superconductor blog: Grok 4.6 joins the Pantheon (August 13, 2026) (checked November 4, 2026)
- Superconductor docs: Agent credentials (checked November 4, 2026)
- Superconductor blog: Kimi K3 benchmark (August 3, 2026) (checked November 5, 2026)
- Claude Code docs: Permission modes (checked November 5, 2026)
- Claude Code docs: Sandboxing (checked November 5, 2026)
- Claude Code docs: Memory and AGENTS.md (checked November 5, 2026)
- Claude Code docs: Desktop (checked November 5, 2026)
- Codex docs: Remote connections (checked November 5, 2026)
- Codex docs: Enterprise admin FAQ (checked November 5, 2026)
- Atlassian: Connect ChatGPT and Codex to Jira and Confluence (checked November 5, 2026)
- OpenAI Trust Portal (checked November 5, 2026)
- TechCrunch: DeepSeek, the countries and agencies that have banned the AI company's tech (checked November 5, 2026)
- The Multiplayer AI Manifesto: Stay provider-agnostic (checked November 5, 2026)
- Garante Privacy: Il Garante blocca ChatGPT (March 31, 2023) (checked November 5, 2026)
- CBS News: Trump orders federal agencies to stop using Anthropic's AI technology (February 27, 2026) (checked November 5, 2026)
Products and prices change. If something here is wrong or out of date, email team@superconductor.com and we will correct it.